Function xsum_carry_propagate

xs/sources/xsum.c:180–462

   2^XSUM_LOW_MANTISSA_BITS - 1. */

static NOINLINE int xsum_carry_propagate (xsum_small_accumulator *restrict sacc)
{
  int i, u, uix;

  if (xsum_debug) c_printf("\nCARRY PROPAGATING IN SMALL ACCUMULATOR\n");

  /* Set u to the index of the uppermost non-zero (for now) chunk, or
     return with value 0 if there is none. */

# if OPT_CARRY

  { u = XSUM_SCHUNKS-1;
    switch (XSUM_SCHUNKS & 0x3)   /* get u to be a multiple of 4 minus one  */
    {
      case 3: if (sacc->chunk[u] != 0)
              { goto found2;
              }
              u -= 1;                            /* XSUM_SCHUNKS is a */
      		  mxFallThrough;
      case 2: if (sacc->chunk[u] != 0)           /* constant, so the  */
              { goto found2;                     /* compiler will do  */
              }                                  /* simple code here  */
              u -= 1;
      		  mxFallThrough;
      case 1: if (sacc->chunk[u] != 0)
              { goto found2;
              }
              u -= 1;
      		  mxFallThrough;
      case 0: ;
    }

    do  /* here, u should be a multiple of 4 minus one, and at least 3 */
    {
#     if USE_SIMD && __AVX__
      { __m256i ch;
        ch = _mm256_loadu_si256 ((__m256i *)(sacc->chunk+u-3));
        if (!_mm256_testz_si256(ch,ch))
        { goto found;
        }
        u -= 4;
        if (u < 0)  /* never actually happens, because value of XSUM_SCHUNKS */
        { break;    /*   is such that u < 0 occurs at end of do loop instead */
        }         
        ch = _mm256_loadu_si256 ((__m256i *)(sacc->chunk+u-3));
        if (!_mm256_testz_si256(ch,ch))
        { goto found;
        }
        u -= 4;
      }
#     else
      { if (sacc->chunk[u] | sacc->chunk[u-1]
          | sacc->chunk[u-2] | sacc->chunk[u-3])
        { goto found;
        }
        u -= 4;
      }
#     endif